Danish supercomputer predicts who will develop deadly cancer

Abdominal pain, weight loss, diabetes are symptoms of pancreatic cancer, which is a relatively rare but often deadly form of cancer.

Each year, around it accounts for around 3 percent of all cancer diagnosis, and the number is growing considerably.

This cancer is difficult to predict, since symptoms are common and overlap with a number of less dangerous diseases. Therefore, only around 10 percent of patients are alive five years after being diagnosed with pancreatic cancer.

However, a new international study based on Danish health data and led from  the University of Copenhagen shows that it is possible to better predict who is at risk of developing the disease.

“If we are able to detect the cancer early, patients can benefit from surgery extending the patient’s life. We have trained a supercomputer to scan health data from millions of Danes and identify the factors that increase the risk of pancreatic cancer. Using our algorithm, we are able to reach the right people earlier,” says leading author of the new study Søren Brunak, who is a professor at the Novo Nordisk Foundation Center for Protein Research, University of Copenhagen.

Examined millions of people’s medical history

The researchers behind the new study examined more than 6.5 million Danes’ medical history going back to 1977 and asked the computer to identify patterns characterizing the people who were diagnosed with pancreatic cancer later in life.

What is unique about this study is the fact that the researchers went on to train the computer using health data from the US. Based on data from around three million Americans, the computer identified similar patterns characterizing pancreatic cancer patients despite many differences in the underlying health data.

“What we have done is to take your entire medical history and sorted all your diagnoses into chronological order. We have then asked the computer to determine which diagnosis patterns may lead to pancreatic cancer later in life and compared the two different algorithms across two countries to strengthen our findings,” Søren Brunak explains and adds:

“But aside from that, we did not tell the computer what to look for. This is what is called non-hypothesis-driven research; where you may discover features you did not specifically select.”

Supercomputer identified new risk factors

The study, which has been published in the renowned journal Nature Medicine, has been a collaboration between researchers at the University of Copenhagen and Harvard University, respectively.

According to the two first authors from the University of Copenhagen, Davide Placido and Jessica Xin Hjaltelin, the factors identified by the supercomputer include combinations of the 16 known pancreatic cancer risk factors previously identified.

“The prediction produced by our algorithm is based on combinations of risk factors such as type 2 diabetes and chronic pancreatitis. Risk factors are not necessarily the same for all patients. Aside from these known factors, the model also pointed at digestive disorders and symptoms such as gallstone disease, gastritis, pyrosis and heartburn,” says PhD Student Davide Placido and Assistant Professor Jessica Xin Hjaltelin from the Novo Nordisk Foundation Center for Protein Research at the University of Copenhagen.

“The study does not say anything about how to treat the disease, but we hope that doctors and the healthcare systems may use it to screen patients in the future.”

Read the full study here: ’A deep learning algorithm to predict risk of pancreatic cancer from disease trajectories’.